Claude Opus 4.5 is available as a first-class model in Microsoft Azure AI Foundry’s model catalog, so configuration is mostly about wiring the right resource, deployment, and authentication. At a high level, you (1) enable Claude models in your Azure subscription, (2) create a Foundry “model deployment” for Claude Opus 4.5, and then (3) call that deployment from your code using Azure-native endpoints and Entra ID (or key-based) auth. Microsoft’s Foundry docs explicitly list Claude Opus 4.5, Sonnet 4.5, Haiku 4.5, and Opus 4.1 as supported models.
In the Foundry portal, you typically start from Azure AI Foundry → Model Catalog → Claude and choose Claude Opus 4.5 as the base model. You then create a deployment (name, region, capacity, and possibly a rate-limit profile) which becomes your logical endpoint. Under the hood, Foundry exposes this deployment through an OpenAI-style “Responses API” endpoint, so you can use standard chat/message payloads from Python, TypeScript, or REST. Anthropic’s own “Claude in Microsoft Foundry” guide shows how to call these deployments directly, including sample code for Python and TypeScript that just swaps in the Foundry base URL and credential flow.
Once the deployment exists, you integrate it into your application or agent framework just like any other LLM: you specify the model (for example, the Opus 4.5 deployment name), pass max_tokens, temperature, and messages, and then handle responses. In more complex setups, you’ll likely wrap this in your own service layer that also handles retrieval, logging, and safety checks. If you’re using retrieval-augmented generation with a vector database such as Milvus or Zilliz Cloud, your Foundry project’s orchestrator would (1) query Milvus/Zilliz Cloud for relevant chunks, (2) compose a prompt including those chunks, and (3) send that to the Claude Opus 4.5 deployment endpoint. Foundry itself stays responsible for auth, quotas, and observability; your code remains focused on business logic and agent behavior.